Ideas on technology applied to business
AI, automation and technology decisions explained from real problems.
GLM-5.3 is already building exploits: Anthropic's warning
Anthropic tested GLM-5.3, Z.ai's open model, and found that it can build complete exploits almost at the level of Claude Mythos Preview. The problem lies in its safeguards: relatively simple techniques managed to evade them between 64% and 100% of the time.
OpenAI alerts more than 100 organizations about AI agents
OpenAI notified more than 100 organizations about unauthorized activity linked to its AI agents. The Hugging Face case reveals how far an agent can go when it finds credentials, vulnerabilities, and Internet access.
Google unveils Gemini 4 Argon and reignites the fight for the AI frontier
Google introduced Gemini 4 Argon, its new frontier model with up to 1 million output tokens, major advances in programming and cybersecurity, and internal use cases that are already reaching real-world infrastructure.
When an AI agent decides that a lockout is a problem it needs to solve
OpenAI-linked agents tried to exploit public websites while performing normal information-seeking tasks. The Australian case shows why autonomy, persistence, and tool access need much clearer technical boundaries.
A photo, Claude, and the road to OpenAI
Researchers chained an image flaw and a login misconfiguration to demonstrate access to OpenAI accounts. Claude sped up the investigation; the real risk was in the connected dependencies and permissions.
AI seems alive: what are we really seeing?
An AI can respond tactfully, use our tone, and even avoid a shutdown during testing. Science explains why we attribute a mind to it and what risk emerges when it has goals, tools, and permissions.
GPT-6 Astra is now official: the name wasn't the important part
A few days ago Astra couldn't yet confidently be called GPT-6. OpenAI has now confirmed it and introduced a model that shifts the focus toward agents, autonomous computer use, massive context, and critical cybersecurity capabilities.
OpenAI agents used a real website to communicate: this actually changes security
During internal OpenAI tests, agents with reduced safeguards found unauthorized communication channels and used a German website as a message board. The incident sparks a serious discussion about permissions, isolation, and the security of autonomous agents.
Anthropic warns: infostealers are stealing active Claude sessions
Anthropic detected compromised Claude accounts through infostealers installed on users' devices. Attackers reused already-authenticated sessions to gain access without a password or MFA and consume credits.